From 096ba84f00636bf3e317e1f8fac42a4c3f3834ce Mon Sep 17 00:00:00 2001
From: yucheng <yucheng@berri.ai>
Date: Wed, 23 Sep 2026 08:00:26 +0000
Subject: [PATCH] fix(batches): map anthropic batch lines through the anthropic
 message transformer

Anthropic result lines carry a message-shaped body, so ModelResponse(**body) produced empty choices and a usage key that blocked the batch-rate usage attach. Extract the existing AnthropicResponse -> ModelResponse conversion used for /v1/messages logging into anthropic_message_to_model_response and reuse it for anthropic line items.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/batches/batch_line_item_logging.py    | 10 ++++--
 litellm/litellm_core_utils/litellm_logging.py | 33 +++++++++++--------
 .../batches/test_batch_line_item_logging.py   |  6 ++++
 3 files changed, 34 insertions(+), 15 deletions(-)

diff --git a/litellm/batches/batch_line_item_logging.py b/litellm/batches/batch_line_item_logging.py
index 911a4d478ec..b91ef3dccd3 100644
--- a/litellm/batches/batch_line_item_logging.py
+++ b/litellm/batches/batch_line_item_logging.py
@@ -121,11 +121,17 @@ def _line_model(
 _BatchLineResult: TypeAlias = "ModelResponse | EmbeddingResponse | ResponsesAPIResponse"
 
 
-def _line_result(call_type: str, response_body: Mapping[str, object]) -> _BatchLineResult:
+def _line_result(
+    call_type: str, custom_llm_provider: _BatchLineProvider, response_body: Mapping[str, object]
+) -> _BatchLineResult:
     if call_type == "aembedding":
         return EmbeddingResponse(**response_body)  # pyright: ignore[reportArgumentType]  # provider output bodies are dicts expanded as response ctor kwargs
     if call_type == "aresponses":
         return ResponsesAPIResponse(**response_body)  # pyright: ignore[reportArgumentType]  # same as above
+    if custom_llm_provider == "anthropic":
+        from litellm.litellm_core_utils.litellm_logging import anthropic_message_to_model_response
+
+        return anthropic_message_to_model_response(response_body, None)
     return ModelResponse(**response_body)  # pyright: ignore[reportArgumentType]  # same as above
 
 
@@ -235,7 +241,7 @@ async def _emit_line_event(
 
     stats: Final = _safe_output_line_stats(entry, custom_llm_provider, model_name, model_info)
     try:
-        result: Final = _line_result(call_type, response_body)
+        result: Final = _line_result(call_type, custom_llm_provider, response_body)
     except Exception:  # noqa: BLE001  # one unparseable line must not drop the rest of the batch's line events
         verbose_logger.warning(
             "batch output line could not be reconstructed as a %s response, skipping it. custom_id=%s",
diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 06c8218eea5..11b1d4725f7 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -535,6 +535,24 @@ def _provider_response_id(source: object) -> str | None:
     return candidate if isinstance(candidate, str) and candidate else None
 
 
+def anthropic_message_to_model_response(result: Mapping[str, object], speed: str | None) -> ModelResponse:
+    import httpx
+
+    from litellm.types.llms.anthropic import AnthropicResponse
+
+    pydantic_result: Final = AnthropicResponse.model_validate(result)
+    return litellm.AnthropicConfig().transform_parsed_response(
+        completion_response=pydantic_result.model_dump(),
+        raw_response=httpx.Response(
+            status_code=200,
+            headers={},
+        ),
+        model_response=litellm.ModelResponse(id=_provider_response_id(result)),
+        json_mode=None,
+        speed=speed,
+    )
+
+
 def mask_api_base_credentials(api_base: str) -> str:
     if "key=" not in api_base:
         return api_base
@@ -4237,19 +4255,8 @@ class Logging(LiteLLMLoggingBaseClass):
                 litellm_params={},
             )
         else:
-            from litellm.types.llms.anthropic import AnthropicResponse
-
-            pydantic_result: Final = AnthropicResponse.model_validate(result)
-            import httpx
-
-            result = litellm.AnthropicConfig().transform_parsed_response(
-                completion_response=pydantic_result.model_dump(),
-                raw_response=httpx.Response(
-                    status_code=200,
-                    headers={},
-                ),
-                model_response=litellm.ModelResponse(id=provider_response_id),
-                json_mode=None,
+            result = anthropic_message_to_model_response(
+                cast(Mapping[str, object], result),
                 speed=self.optional_params.get("speed") if self.optional_params else None,
             )
         return result
diff --git a/tests/test_litellm/batches/test_batch_line_item_logging.py b/tests/test_litellm/batches/test_batch_line_item_logging.py
index fe03e348ce0..62c6bd359f2 100644
--- a/tests/test_litellm/batches/test_batch_line_item_logging.py
+++ b/tests/test_litellm/batches/test_batch_line_item_logging.py
@@ -368,6 +368,8 @@ ANTHROPIC_OUTPUT_JSONL = json.dumps(
                 "role": "assistant",
                 "content": [{"type": "text", "text": "hello b2"}],
                 "model": "claude-3",
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
                 "usage": {"input_tokens": 1, "output_tokens": 2},
             },
         },
@@ -482,3 +484,7 @@ async def test_line_items_anthropic_shapes(recorder):
     line = next(e for e in recorder.success_events if _hidden(e).get("batch_custom_id") == "b2")
     assert _hidden(line)["batch_line_status_code"] == 200
     assert line["litellm_params"]["batch_parent_id"] == batch.id
+    payload = _payload(line)
+    assert payload["response"]["choices"][0]["message"]["content"] == "hello b2"
+    assert payload["prompt_tokens"] == 1
+    assert payload["completion_tokens"] == 2